Worker spout failure behaviour

[Simon Cooper <si...@featurespace.co.uk>]

What’s the defined failure behaviour for storm? In my testing, if a partitioned trident spout (or the worker running a spout) crashes, then the next transaction is run with a null metadata. This is highly undesirable behaviour, as there is no way for the spout to know where it is, or where the emitter got to before the worker crashed. So there’s no way to work out if the emitter has missed some of the backing data (say, if it’s emitting from a kafka queue, and hadn’t caught up with realtime).

Is this behaviour intentional? When a spout crashes, the next transaction is run with null metadata?

Thanks,
SimonC